OpenPose
An open-source human keypoint detection framework proposed in 2017 by the Perceptual Computing Lab at Carnegie Mellon University. Based on Part Affinity Fields (PAF), it performs real-time multi-person keypoint detection and outputs body, hand, and facial skeletons. It is widely used in controllable AI video generation.
Core Facts
Timeline (last 90 days)
DensePose and OpenPose are representative early works responsible for extracting skeletal keypoints.
The LoRA-based pose control has relatively softer constraints and relies more on prompt cooperation, but requires no additional preprocessing steps (such as extracting skeleton maps with OpenPose).
DensePose and OpenPose handle pose estimation, while models such as Viggle and AnimateAnyone integrate the motion transfer pipeline end-to-end
When character body proportions are inconsistent, ControlNet based on OpenPose skeletons has difficulty producing ideal results.
Controlling poses with ControlNet requires first converting the control video into OpenPose skeletons, which adds extra processing time.
OpenPose can detect human body keypoints and represent the skeleton as a colored line diagram, with keypoint coordinates expressed as absolute pixel positions.
Pose estimation models (such as MediaPipe or OpenPose) can identify the spatial coordinates of key human body nodes such as shoulders, elbows, and knees.
MediaPipe uses a two-stage detector and can achieve over 30 fps on a CPU, compared to OpenPose which requires GPU acceleration.
Human pose estimation technology can extract 3D skeletal motion sequences from 2D video, with mainstream solutions including OpenPose, MediaPipe, and MotionBERT.
骨骼关键点检测（如OpenPose、MediaPipe、DWPose）通过定位人体17至25个关节点描述整体姿态，是稀疏但语义化的表示方法
All Facts (12)
OpenPose was proposed around 2017 by the Perceptual Computing Lab at Carnegie Mellon University and is the first open-source framework capable of real-time multi-person body keypoint detection.
75%Unverified姿势识别功能依赖OpenPose等人体姿态估计算法，从参考图中提取18个人体关键点（头、肩、肘、腕、髋、膝、踝等）生成骨骼图
85%UnverifiedDensePose and OpenPose are representative early works responsible for extracting skeletal keypoints.
50%UnverifiedThe LoRA-based pose control has relatively softer constraints and relies more on prompt cooperation, but requires no additional preprocessing steps (such as extracting skeleton maps with OpenPose).
50%UnverifiedDensePose and OpenPose handle pose estimation, while models such as Viggle and AnimateAnyone integrate the motion transfer pipeline end-to-end
50%UnverifiedControlling poses with ControlNet requires first converting the control video into OpenPose skeletons, which adds extra processing time.
50%UnverifiedWhen character body proportions are inconsistent, ControlNet based on OpenPose skeletons has difficulty producing ideal results.
50%UnverifiedOpenPose can detect human body keypoints and represent the skeleton as a colored line diagram, with keypoint coordinates expressed as absolute pixel positions.
50%UnverifiedPose estimation models (such as MediaPipe or OpenPose) can identify the spatial coordinates of key human body nodes such as shoulders, elbows, and knees.
50%UnverifiedMediaPipe uses a two-stage detector and can achieve over 30 fps on a CPU, compared to OpenPose which requires GPU acceleration.
50%UnverifiedHuman pose estimation technology can extract 3D skeletal motion sequences from 2D video, with mainstream solutions including OpenPose, MediaPipe, and MotionBERT.
50%Unverified骨骼关键点检测（如OpenPose、MediaPipe、DWPose）通过定位人体17至25个关节点描述整体姿态，是稀疏但语义化的表示方法
50%